Overview
- Several groups published open-source GraphRAG systems on July 14–15 that split extraction from consolidation and ship code and models, including RAGU with a compact 7B extractor model (Meno-Lite-0.1) and an SVD-based tree builder (SVD-RAG) that claims large speed and token-cost savings.
- FAIR GraphRAG embeds FAIR Digital Objects as graph nodes for biomedical RNA‑seq data and reports gains in question-answer accuracy, coverage, and explainability by pairing metadata-rich nodes with LLM-assisted schema and extraction.
- QIMG-7, a new multimodal benchmark, shows that corrupted images and text can severely degrade RAG answers and that source-aware trust resolution (SATR) recovers much of the lost accuracy by selecting or falling back to more reliable candidate sources.
- Agentic approaches advanced with GRASP, which trains RL policies to pick search granularity (semantic search, keywords, paragraph reads) and improves multi-hop recall, and EvoGraph-R1, which treats the knowledge graph as a self-evolving environment reshaped by agent actions to refine evidence and structure.
- The combined work points to practical trade-offs for builders: cheaper deterministic summarization and smaller in-pipeline models can lower cost and speed, but multimodal corruption and ideological bias in retrieved corpora raise new reliability risks that require benchmarks, selective trust rules, and further peer review.